It looks like the majority of what I say below is not related to the fate of the "bit". The push to take the bit was strong with this one, and me... can't we deploy more of what we already got in places where it matters?
... so: A) PLEA: From 10 years now, of me working on bufferbloat, working on real end-user and wifi traffic and real networks.... I would like folk here to stop benchmarking two flows that run for a long time and in one direction only... and thus exclusively in tcp congestion avoidance mode. Please. just. stop. Real traffic looks nothing like that. The internet looks nothing like that. The netops folk I know just roll their eyes up at benchmarks like this that prove nothing and tell me to go to ripe meetings instead. When y'all talk about "not looking foolish for not mandating ecn now", you've already lost that audience with benchmarks like these. Sure, setup a background flow(s) like that, but then hit the result with a mix of far more normal traffic? Please? networks are never used unidirectionally and both directions congesting is frequent. To illustrate that problem... I have a really robust benchmark that we have used throughout the bufferbloat project that I would like everyone to run in their environments, the flent "rrul" test. Everybody on both sides has big enough testbeds setup that a few hours spent on doing that - and please add in asymmetric networks especially - and perusing the results ought to be enlightening to everyone as to the kind of problems real people have, on real networks. Can the L4S and SCE folk run the rrul test some day soon? Please? I rather liked this benchmark that tested another traffic mix, ( https://www.cablelabs.com/wp-content/uploads/2014/06/DOCSIS-AQM_May2014.pdf ) although it had many flaws (like not doing dns lookups), I wish it could be dusted off and used to compare this new fangled ecn enabled stuff with the kind of results you can merely get with packet loss and rtt awareness. It would be so great to be able to directly compare all these new algorithms against this benchmark. Adding in a non ecn'd udp based routing protocol on heavily oversubscribed 100mbit link is also enlightening. I'd rather like to see that benchmark improved for a more modernized home traffic mix where it is projected there may be 30 devices on the network on average, in a few years. If there is any one thing y'all can do to reduce my blood pressure and keep me engaged here whilst you debate the end of the internet as I understand it, it would be to run the rrul test as part of all your benchmarks. thank you. B) Stuart Cheshire regaled us with several anecdotes - one concerning his problems with comcast's 1Gbit/35mbit service being unusable, under load, for videoconferencing. This is true. The overbuffering at the CMTSes still, has to be seen to be believed, at all rates. At lower rates it's possible to shape this, with another device (which is what the entire SQM deployment does in self defense and why cake has a specific docsis ingress mode), but it is cpu intensive and requires x86 hardware to do well at rates above 500Mbits, presently. So I wish CMTS makers (Arris and Cisco) were in this room. are they? (Stuart, if you'd like a box that can make your comcast link pleasurable under all workloads, whenever you get back to los gatos, I've got a few lying around. Was so happy to get a few ietfers this past week to apply what's off the shelf for end users today. :) C) I am glad bob said the L4S is finally looking at asymmetric networks, and starting to tackle ack-filtering and accecn issues there. But... I would have *started there*. Asymmetric access is the predominate form of all edge technologies. I would love to see flent rrul test results for 1gig/35mbit, 100/10, 200/10 services, in particular. (from SCE also!). "lifeline" service (11/2) would be good to have results on. It would be especially good to have baseline comparison data from the measured, current deployment of the CMTSes at these rates, to start with, with no queue management in play, then pie on the uplink, then fq_codel on the uplink, and then this ecn stuff, and so on. D) The two CPE makers in the room have dismissed both fq and sce as being too difficult to implement. They did say that dualpi was actually implemented in software, not hardware. I would certainly like them to benchmark what they plan to offer in L4S vs what is already available in the edgerouter X, as one low end example among thousands. I also have to note, at higher speeds, all the buffering moves into the wifi and the results are currently ugly. I imagine they are exploring how to fix their wifi stacks also? I wish more folk were using RVR + latency benchmarks like this one: http://flent-newark.bufferbloat.net/~d/Airtime%20based%20queue%20limit%20for%20FQ_CoDel%20in%20wireless%20interface.pdf Same goes for the LTE folk. E) Andrew mcgregor mentioned how great it would be for a closeted musician to be able to play in real time with someone across town. that has been my goal for nearly 30 years now!! And although I rather enjoyed his participation in my last talk on the subject ( https://blog.apnic.net/2020/01/22/bufferbloat-may-be-solved-but-its-not-over-yet/ ) conflating a need for ecn and l4s signalling for low latency audio applications with what I actually said in that talk, kind of hurt. I achieved "my 2ms fiber based guitarist to fiber based drummer dream" 4+ years back with fq_codel and diffserv, no ecn required, no changes to the specs, no mandating packets be undroppable" and would like to rip the opus codec out of that mix one day. F) I agree with jana that changing the definition of RFC3168 to suit the RED algorithm (which is not pi or anything fancy) often present in network switches, today to suit dctcp, works. But you should say "configuring red to have l4s marking style" and document that. Sometimes I try to point out many switches have a form of DRR in them, and it's helpful to use that in conjunction with whatever diffserv markings you trust in your network. To this day I wish someone would publish how much they use DCTCP style signalling on a dc network relative to their other traffic. To this day I keep hoping that someone will publish a suitable set of RED parameters for a wide variety of switches and routers - for the most common switches and ethernet chips, for correct DCTCP usage. Mellonox's example: ( https://community.mellanox.com/s/article/howto-configure-ecn-on-mellanox-ethernet-switches--spectrum-x ) is not dctcp specific. many switches have a form of DRR in them, and it's helpful to use that in conjunction with whatever diffserv markings you trust in your network, and, as per the above example, segregate two red queues that way. From what I see above there is no way to differentiate ECT(0) from ECT(1) in that switch. (?) I do keep trying to point out the size of the end user ecn enabled deployment, starting with the data I have from free.fr. Are we building a network for AIs or people? G) Jana also made a point about 2 queues "being enough" (I might be mis-remembering the exact point). Mellonoxes ethernet chips at 10Gig expose 64 hardware queues, some new intel hardware exposes 2000+. How do these queues work relative to these algorithms? We have generally found hw mq to be far less of a benefit than the manufacturers think, especially as regard to lower latency or reduced cpu usage (as cache crossing is a bear). There is a lot of software work in this area left to be done, however they are needed to match queues to cpus (and tenants) Until sch_pie gained timestamping support recently, the rate estimator did not work correctly in a hw mq environment. Haven't looked over dualpi in this respect. -- Make Music, Not War Dave Täht CTO, TekLibre, LLC http://www.teklibre.com Tel: 1-831-435-0729 _______________________________________________ Bloat mailing list Bloat@lists.bufferbloat.net https://lists.bufferbloat.net/listinfo/bloat
